What the study found
A foundation model for tabular data, called TabPFN, produced crop-yield forecasts for maize, soybeans, and sunflowers in South Africa with accuracy comparable to several machine-learning models and better than baseline approaches.
Why the authors say this matters
The authors say accurate and timely crop yield forecasts are crucial for decisions about resource allocation, market strategies, and food security interventions, especially in food-insecure regions. They conclude that TabPFN’s faster tuning time and lower need for feature engineering make it practically useful.
What the researchers tested
The researchers applied TabPFN, a foundation model for small- to medium-sized tabular data, to sub-national yield forecasting in South Africa. They used dekadal Earth Observation data (10-day measurements of Fraction of Absorbed Photosynthetically Active Radiation, or FAPAR, and soil moisture) and gridded weather data (air temperature, precipitation, and radiation), aggregated at a monthly scale by region. They benchmarked TabPFN against six machine-learning models and two baseline models using leave-one-year-out cross-validation.
What worked and what didn't
TabPFN and the machine-learning models had comparable accuracy overall, and both outperformed the baseline approaches. For maize at the national level, the best-performing machine-learning model achieved 6.8% rRMSEp and an R² of 0.91, while TabPFN achieved 8.8% rRMSEp and an R² of 0.86. The abstract reports that TabPFN was useful because it tuned faster and required less feature engineering.
What to keep in mind
The study summary does not describe limitations beyond the scope of the experiment. The results are limited to sub-national yield forecasting in South Africa for maize, soybeans, and sunflowers, using the data and validation setup described.
Key points
- TabPFN forecasted crop yields with accuracy similar to several machine-learning models and better than baseline models.
- The study focused on sub-national yield forecasting for maize, soybeans, and sunflowers in South Africa.
- Inputs included satellite-based Earth Observation data and gridded weather data aggregated by region and month.
- Using leave-one-year-out cross-validation, TabPFN was tested against six machine-learning models and two baseline models.
- For maize at the national level, TabPFN reached 8.8% rRMSEp and an R² of 0.86.
Disclosure
- Research title:
- TabPFN matched standard models for crop yield forecasting
- Authors:
- Filip Sabo, Michele Meroni, María Piles, Martin Claverie, Fanie Ferreira, Elna Van Den Berg, Francesco Collivignarelli, Felix Rembold
- Institutions:
- Universitat de València, European Commission, Joint Research Centre, Oblikue Consulting (Spain), Geoterra Image (South Africa)
- Publication date:
- 2026-04-24
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with peer-reviewed research without reading academic papers — one filtered digest, every Friday.